Proceedings of the ACL-HLT 2011 Student Session, pages 6–11, Portland, OR, USA 19-24 June 2011. c 2011 Association for Computational Linguistics Sentence Ordering Driven by Local and Global Coherence for Summary Generation Renxian Zhang Department of Computing The Hong Kong Polytechnic University csrzhang@comp.polyu.edu.hk Abstract In summarization, sentence ordering is conducted to enhance summary readability by accommodating text coherence. We propose a grouping-based ordering framework that integrates local and global coherence concerns. Summary sentences are grouped before ordering is applied on two levels: group-level and sentence-level. Different algorithms for grouping and ordering are discussed. The preliminary results on single-document news datasets demonstrate the advantage of our method over a widely accepted method. 1 Introduction and Background The canonical pipeline of text summarization consists of topic identification, interpretation, and summary generation (Hovy, 2005). In the simple case of extraction, topic identification and interpretation are conflated to sentence selection and concerned with summary informativeness. In comparison, summary generation addresses summary readability and a frequently discussed generation technique is sentence ordering. It is implicitly or explicitly stated that sentence ordering for summarization is primarily driven by coherence. For example, Barzilay et al. (2002) use lexical cohesion information to model local coherence. A statistical model by Lapata (2003) considers both lexical and syntactic features in calculating local coherence. More globally biased is Barzilay and Lee’s (2004) HMM-based content model, which models global coherence with word distribution patterns. Whilst the above models treat coherence as lexical or topical relations, Barzilay and Lapata (2005, 2008) explicitly model local coherence with an entity grid model trained for optimal syntactic role transitions of entities. Although coherence in those works is modeled in the guise of “lexical cohesion”, “topic closeness”, “content relatedness”, etc., few published works simultaneously accommodate coherence on the two levels: local coherence and global coherence, both of which are intriguing topics in text linguistics and psychology. For sentences, local coherence means the well- connectedness between adjacent sentences through lexical cohesion (Halliday and Hasan, 1976) or entity repetition (Grosz et al., 1995) and global coherence is the discourse-level relation connecting remote sentences (Mann and Thompson, 1995; Kehler, 2002). An abundance of psychological evidences show that coherence on both levels is manifested in text comprehension (Tapiero, 2007). Accordingly, an apt sentence ordering scheme should be driven by such concerns. We also note that as sentence ordering is usually discussed only in the context of multi-document summarization, factors other than coherence are also considered, such as time and source sentence position in Bollegala et al.’s (2006) “agglomerative ordering” approach. But it remains an open question whether sentence ordering is non-trivial for single-document summarization, as it has long been recognized as an actual strategy taken by human summarizers (Jing, 1998; Jing and McKeown, 2000) and acknowledged early in work on sentence ordering for multi-document summarization (Barzilay et al., 2002). In this paper, we outline a grouping-based sentence ordering framework that is driven by the concern of local and global coherence. Summary sentences are grouped according to their conceptual relatedness before being ordered on two levels: group-level ordering and sentence-level ordering, which capture global coherence and local coherence in an integrated model. As a preliminary study, we applied the framework to single- 6 . 6–11, Portland, OR, USA 19-24 June 2011. c 2011 Association for Computational Linguistics Sentence Ordering Driven by Local and Global Coherence for Summary. relatedness before being ordered on two levels: group-level ordering and sentence-level ordering, which capture global coherence and local coherence in